[Yao Xin: Evolutionary computation may be the next hot spot for artificial intelligence] Evolutionary computing is often used in intelligent optimization and machine learning, but this type of machine learning is not very similar to the machine learning that everybody often says is deep learning. Evolutionary computation should have a great deal of use in the integration of brain and body design. Evolutionary computation may be the next hot spot for artificial intelligence. This article is sorted out from the 2018 Shenzhen International Robotics and Intelligent Systems Academician Forum on the IEEEFellow Southern University of Science and Technology Professor Yao Xin named "Why research evolutionary computing?" "" speech.
IEEE Fellow Professor, Department of Computer Science and Engineering, South University of Science and Technology, Professor Yao Xin
"Why research evolutionary computing? 》
We may not know much about the Department of Computer Science and Engineering of the Southern University of Science and Technology. The department was established in August 2016. In 2017, we had the first batch of undergraduates officially recognized by the country. Last year, we recruited 19 people. A master student and 21 doctoral students, and after a year and a half, we now have 19 teachers and we plan to reach 55 in the future.
The research fields of the Department of Computer Science and Engineering of the Southern University of Science and Technology are divided into five major areas, including artificial intelligence, data science, theory, systems and networks, and cognitive and autonomous systems. There are 5 teachers in the artificial intelligence team. I am one of them. The other teachers are from different places and backgrounds are not exactly the same, but they are all related to computational intelligence and neural evolution. In addition, we also have some doctors from all over the world. We are mainly engaged in a lot of machine learning, optimization and research work between them. It is not comprehensive to opt for optimisation of light learning. The purpose of learning is to make decisions, so learning and optimization need to be combined. There are many aspects of optimization considerations, such as multi-objective optimization, dynamic optimization, and optimization in uncertain environments. Machine learning is more about machine learning, online learning of data streams, and unbalanced learning. Another subject of our research team is cognitive and autonomous systems. There are hardware and software. Hardware is drones and group robots. Software is software robots.
Why study evolutionary calculations?
First of all, let's take a look at what is evolutionary computing? I don't know how many people are writing programs for themselves this year. Writing a program is like eating stinky tofu. It either likes or hates it. How do you feel if you particularly like or hate writing programs? Even if the current computer or robot is clever enough to such an extent, you will usually be very hard on the keyboard. Why do you knock on the keyboard? Whatever program you write is not a comma or parenthesis. Editing is always wrong. Everyone has written a program to know that some spaces are not the same as the compilation, the total error of the compiler, you think this is very distressed. You say that AI has done so far. People seldom make the same mistake twice. When you come across this type of person, this kind of person is incurable. However, the computer is sometimes very troublesome. Are you saying that there is not a comma? , But in fact the computer does not understand, so the computer is very inflexible and very fragile.
In addition, the standards and capabilities are relatively poor. Everyone now has a few people using a computer ten years ago? No, the computer is changed every three years. As people get smarter as they grow older, robots and computers must be replaced in a few years. The path of machine development is not quite the same as that of human development. Natural systems can be considered as a set of computing systems. In contrast, some features of natural systems are indeed very good, but computer science cannot be done for the time being. For example, self-recovery, self-adaptation and so on. There are many places in the natural world that are worth learning for computer science. Therefore, finding inspiration from nature is not unique to computer science. The engineering community often finds inspiration from nature, such as aircraft, from bird's flight to dual fixed wing's. The aircraft and the propeller plane are all in search of inspiration from the birds' flight.
Why learn from nature? The problem solving method of the natural system is complementary to the problem solving method of the computer, and the solution is usually relatively simple and not particularly complicated. This is why it is necessary to study evolutionary computing. In fact, this kind of research is not simple, and not only for the evolutionary computation but also for the in-depth application of neural networks. The artificial neural network is also inspired and influenced by the brain. This evolutionary computation or evolutionary algorithm is through the evolution of biology. Come from. All the living things in the world are evolved. There are certain rules behind them. Even if one thousandth or one thousandth is found in the law, it may bring different ideas to computer design. This is to do evolutionary calculations. A starting point.
What is Evolutionary Computing?
Here are four examples to illustrate.
In the first example, someone who likes machine learning calls this thing machine learning. In fact, it is a data-driven model. Give you a lot of experimental data, and then ask you what is the abstract model behind the data. You help me find out, that is, look for the law from the data. Taking the design of aluminum alloy materials as an example, computer science is a very interesting field. We look at the design of aluminum alloy materials. It used to be modeling. Now we try to use evolutionary algorithms to minimize the time in the laboratory. Assume that we know that the synthesis of this aluminum alloy, you can do preliminary experiments, in the laboratory to pull it with a certain force and pressure it, after the failure to see the deformation of the aluminum ingot, the experimental experts can write a set of equations There are four equation groups here. This example has six material constants. This material constant is a value that we really want to know when we actually design a new material. We don't actually know that this value is a material material scientist, and it's quite a mathematical computer scientist. For variables. I will now give you the equations, give you experimental data, and say what the behavior of this material is in the experiment. Can you find the constant of the material? This is a bit like solving equations. What is the relationship between solving this equation and doing computer science? What is the relationship with evolutionary algorithms? This equation has no analytical solution. It can only be solved numerically. One way is to turn it into an optimization method. When looking for a numerical solution, look at the difference between the left and right sides of the equation. Add the difference between the two sides of the equation to zero. found it. How to find it? I'm looking for a bit of a software package. I'm going to buy one of the best packages in the world. Out of the Oxford software package, initialize the initial six variables you're looking for, and then lose the evolution. You find yourself on this issue. The result is always the same as the initial value you entered, because the assumptions made in many digital software packages are not valid in practical problems. Under this condition, you can use the evolution algorithm to do the optimization. This optimization is to solve the equation. Find a digital map. The solution found is the best for designing aluminum ingots and aluminum alloys. It is the most accurate numerical constant and material constant.
In the second small example, we talked about optimization, but there are often unwritten assumptions during optimization. The optimization environment and goals that need to be optimized are immutable, but they will change in real life. You give me a goal and indicators to let me do it. I do it halfway to suddenly change this indicator. This is not okay, it can happen in the actual process. For example, in the winter in the north, the road must be salted and frozen. Different countries have different laws. For example, in the United Kingdom, there is a law. A road in two hours must be salted as long as the weather forecast says that the pavement temperature is below 2 degrees. Now there is a specific problem. I have a team. There is a network of roads A. How the team dispatches the cars and allows all road networks to go through within two hours. All kinds of restrictions here make the actual problems not found in operations research or math books. The mathematics says: Suppose the truck's capacity is Q, but the problem we see is that a government, a small team, and 11 cars, and the road conditions can not assume that the car moves forward at an even speed, especially in winter. . You have 11 cars, make a schedule, leave 10 cars in half, and 1 car is broken, but you still have to complete the task, how to complete it dynamically, the mathematical optimization method can not be solved, we still use Evolutionary algorithms, algorithms inspired by nature can solve this complexity problem.
The third small example, talk about multi-objective optimization. “More, faster, better, and better†is the slogan of the 1970s. Everyone likes to say it, but it's hard to do it. It's more, faster, better, and provincial. It can't just look at an indicator. It needs several indicators to meet at the same time, as a decision maker needs Selecting various compromises is a typical scenario for multi-objective optimization.
A specific example is the unmanned system. The unmanned system really depends on the software being controlled. Whether you buy a software system or develop a software system yourself, you must demonstrate that the software system is correct. Then you need to do software certification. If you can't prove it, you need to do software testing. All kinds of environments need to be tested. But one key point is that it is impossible to make all the possibilities possible. In the limited resources and limited time, testing each module of the large software system maximizes the test accuracy of the system, and allocates limited people and limited money to the large software module. The goal is the overall software accuracy. The highest, these need to use evolutionary computing methods.
What is the relationship between evolutionary computing and robotics?
Finally, we talk about the relationship between evolutionary computing and robots. In this field, there is a branch called evolutionary robots. The world’s creatures, including the human brain, have evolved. Since the brain of the biological community can evolve, why can't the robot evolve?
One of the benefits of designing a robot with an evolutionary approach is that it can simultaneously design the robot's control system and the robot's shape. Most of the time the research in these two fields is done by people from different disciplines, some doing machine learning, and some doing robotics. But in fact, we should combine the design controller with the group that designed the control form and should not be separated.
For everyone to do a simulation experiment, this is to manually construct a swimming line diagram. In this section of the line chart, each section is the same, except that there is a head on the left, a tail on the right, the middle of each section is the same . I want to do a very simple experiment, design or evolve a line chart, swim, swim from A to B, the sooner the better, but I do not tell this line chart should swim. I don't have any additional information here to tell this line chart how to swim. Just tell you this task. You swim from A to B. You can swim straight and swim as fast as possible. How to control this line chart? Each section of the line graph controls the movement of neurons, which is related to the location of the neurons in the neural network. There are round holes in this line graph, and the neurons are distributed in the physical positions of the circular holes. I evolved a line graph that originally did not understand anything. It was from A to B that evolved the controllers on the neural network. In another experiment, I artificially shifted the line graph on the left side of this line graph to the right, not one's own biased, artificially biased, and the right one's line graph was shifted to the left, not straight. I do this experiment, the same task is to travel from A to B, the same goal, the faster you swim the better, the method is to use evolutionary methods. It is to swim slowly, which swims quickly, and the survival of the fittest is passed on to the next generation and passed on from generation to generation. I want to see how these two sets of experiments come out of neural network control. You will find it very interesting. In the first case, when the line graph has not changed, the shape is still straight at the beginning. The last evolved neural network is the figure in the lower right corner. The red dot is the position of the nerve. symmetry. People don't tell it anything at all. The only feedback is to swim fast and slow. The upper left corner is a line drawing of the initial painting. How to swim is random, and it begins to play in the water. Longer time, after 1190, it found a very fast swimming thread. In the second case, I change the shape of this line graph and use this dotted line to represent it. The structure of the neural network just complements this morphological defect to see if you are left or right.
Samsung Screen Protector,Tempered Glass Samsung,Mobile Tempered Glass,Tempered Glass Screen Protector Iphone Samsung
Shenzhen TUOLI Electronic Technology Co., Ltd. , https://www.szhydrogelprotector.com